1. 1 CSC 427: Data Structures and Algorithm Analysis Fall 2008 Dynamic programming  top-down vs. bottom-up  divide & conquer vs. dynamic programming  examples: Fibonacci sequence, binomial coefficient  caching  example: making change  problem-solving approaches summary

2. 2 Divide and conquer divide and conquer is a top-down approach to problem solving  start with the problem to be solved (i.e., the top)  break that problem down into smaller pieces and solve  continue breaking down until reach base/trivial case (i.e., the bottom) divide and conquer works well when the pieces can be solved independently  e.g., merge sort – sorting each half can be done independently, no overlap what about Fibonacci numbers? 1, 1, 2, 3, 5, 8, 13, 21, … public static int fib(int n) { if (n <= 1) { return 1; } else { return fib(n-1) + fib(n-2); }

3. 3 Top-down vs. bottom-up divide and conquer is a horrible way of finding Fibonacci numbers  the recursive calls are NOT independent; redundencies build up public static int fib(int n) { if (n <= 1) { return 1; } else { return fib(n-1) + fib(n-2); } fib(5) fib(4) + fib(3) fib(3) + fib(2) fib(2) + fib(1).. in this case, a bottom-up solution makes more sense  start at the base cases (the bottom) and work up to the desired number  requires remembering the previous two numbers in the sequence public static int fib(int n) { int prev = 1, current = 1; for (int i = 1; i < n; i++) { int next = prev + current; prev = current; current = next; } return current; }

4. 4 Dynamic programming dynamic programming is a bottom-up approach to solving problems  start with smaller problems and build up to the goal, storing intermediate solutions as needed  applicable to same types of problems as divide & conquer, but bottom-up  usually more effective than top-down if the parts are not completely independent (thus leading to redundancy) example: binomial coefficient C(n, k) is relevant to many problems  the number of ways can you select k lottery balls out of n  the number of birth orders possible in a family of n children where k are sons  the number of acyclic paths connecting 2 corners of an k  (n-k) grid  the coefficient of the x k y n-k term in the polynomial expansion of (x + y) n

5. 5 Example: binary coefficient while easy to define, a binomial coefficient is difficult to compute e.g, 6 number lottery with 49 balls  49!/6!43! 49! = 608,281,864,034,267,560,872,252,163,321,295,376,887,552,831,379,210,240,000,000,000 could try to get fancy by canceling terms from numerator & denominator  can still can end up with individual terms that exceed integer limits a computationally easier approach makes use of the following recursive relationship e.g., to select 6 lottery balls out of 49, partition into: selections that include 1 (must select 5 out of remaining 48) + selections that don't include 1 (must select 6 out of remaining 48)

6. 6 Example: binomial coefficient could use straight divide&conquer to compute based on this relation /** * Calculates n choose k (using divide-and-conquer) * @param n the total number to choose from (n > 0) * @param k the number to choose (0 <= k <= n) * @return n choose k (the binomial coefficient) */ public static int binomial(int n, int k) { if (k == 0 || n == k) { return 1; } else { return binomial(n-1, k-1) + binomial(n-1, k); } however, this will take a long time or exceed memory due to redundant work

7. 7 Dynamic programming solution /** * Calculates n choose k (using dynamic programming) * @param n the total number to choose from (n > 0) * @param k the number to choose (0 <= k <= n) * @return n choose k (the binomial coefficient) */ public static int binomial(int n, int k) { if (n < 2) { return 1; } else { int bin[][] = new int[n+1][n+1]; // CONSTRUCT A TABLE TO STORE for (int r = 0; r <= n; r++) { // COMPUTED VALUES for (int c = 0; c <= r && c <= k; c++) { if (c == 0 || c == r) { bin[r][c] = 1; // ENTER 1 IF BASE CASE } else { // OTHERWISE, USE FORMULA bin[r][c] = bin[r-1][c-1] + bin[r-1][c]; } } } return bin[n][k]; // ANSWER IS AT bin[n][k] } } could instead work bottom-up, filling a table starting with the base cases (when k = 0 and n = k) 0123…k 01 111 211 311 ……… n11

8. 8 Dynamic programming & caching when calculating C(n,k), the entire table must be filled in (up to the nth row and kth column)  working bottom-up from the base cases does not waste any work 0123…k 01 111 211 311 ……… n11 for many problems, this is not the case  solving a problem may require only a subset of smaller problems to be solved  constructing a table and exhaustively working up from the base cases could do lots of wasted work  can dynamic programming still be applied?

9. 9 Example: programming contest problem

10. 10 Divide & conquer approach let getChange(amount, coinList) represent the number of ways to get an amount using the specified list of coins getChange(amount, coinList) = getChange(amount-biggestCoinValue, coinList) // # of ways that use at least + // one of the biggest coin getChange(amount, coinList-biggestCoin) // # of ways that don't // involve the biggest coin e.g., suppose want to get 10¢ using only pennies and nickels getChange(10, [1¢, 5¢]) getChange(5, [1¢, 5¢]) + getChange(10, [1¢]) getChange(0, [1¢, 5¢]) + getChange(5, [1¢]) 1 1 1

11. Divide & conquer solution public class ChangeMaker { private List coins; public ChangeMaker(List coins) { this.coins = coins; } public int getChange(int amount) { return this.getChange(amount, this.coins.size()-1); } private int getChange(int amount, int maxCoinIndex) { if (amount < 0 || maxCoinIndex < 0) { return 0; } else if (amount == 0) { return 1; } else { return this.getChange(amount-this.coins.get(maxCoinIndex), maxCoinIndex) + this.getChange(amount, maxCoinIndex-1); } base case: if amount or max coin index becomes negative, then can't be done base case: if amount is zero, then have made exact change recursive case: count how many ways using a largest coin + how many ways not using a largest coin could implement as a ChangeMaker class  when constructing, specify a sorted list of available coins (why sorted?)  recursive helper method works with a possibly restricted coin list

12. 12 Will this solution work? certainly, it will produce the correct answer -- but, how quickly?  at most 10 coins  worst case: 1 2 3 4 5 6 7 8 9 10  6,292,069 combinations  depending on your CPU, this can take a while  # of combinations will explode if more than 10 coins allowed the problem is duplication of effort getChange(100, 9) getChange(90, 9) + getChange(100, 8) getChange(80, 9) + getChange(90, 8) getChange(91, 8) + getChange(100, 7)................ getChange(80, 6) getChange(80, 6) getChange(80, 6) getChange(80, 6)

13. 13 Caching we could use dynamic programming and solve the problem bottom up  however, consider getChange(100, [1¢, 5¢, 10¢, 25¢])  would we ever need to know getChange(99, [1¢, 5¢, 10¢, 25¢]) ? getChange(98, [1¢, 5¢, 10¢, 25¢]) ? getChange(73, [1¢, 5¢]) ? when exhaustive bottom-up would yield too many wasted cases, dynamic programming can instead utilize top-down with caching  create a table (as in the exhaustive bottom-up approach)  however, fill the table in using a top-down approach that is, execute a top-down divide and conquer solution, but store the solutions to subproblems in the table as they are computed  before recursively solving a new subproblem, first check to see if its solution has already been cached  avoids the duplication of pure top-down  avoids the waste of exhaustive bottom-up (only solves relevant subproblems)

14. 14 ChangeMaker with caching public class ChangeMaker { private List coins; private static final int MAX_AMOUNT = 100; private static final int MAX_COINS = 10; private int[][] remember; public ChangeMaker(List coins) { this.coins = coins; this.remember = new int[ChangeMaker.MAX_AMOUNT+1][ChangeMaker.MAX_COINS]; for (int r = 0; r < ChangeMaker.MAX_AMOUNT+1; r++) { for (int c = 0; c < ChangeMaker.MAX_COINS; c++) { this.remember[r][c] = -1; } public int getChange(int amount) { return this.getChange(amount, this.coins.size()-1); } private int getChange(int amount, int maxCoinIndex) { if (maxCoinIndex < 0 || amount < 0) { return 0; } else if (this.remember[amount][maxCoinIndex] == -1) { if (amount == 0) { this.remember[amount][maxCoinIndex] = 1; } else { this.remember[amount][maxCoinIndex] = this.getChange(amount-this.coins.get(maxCoinIndex), maxCoinIndex) + this.getChange(amount, maxCoinIndex-1); } return this.remember[amount][maxCoinIndex]; } with caching, even the worst case is fast: 6,292,069 combinations as each subproblem is solved, its solution is stored in a table each call to getChange checks the table first before recursing

15. 15 Algorithmic approaches summary divide & conquer: tackles a complex problem by breaking it into smaller pieces, solving each piece, and combining them into an overall solution  often involves recursion, if the pieces are similar in form to the original problem  applicable for any application that can be divided into independent parts dynamic: bottom-up implementation of divide & conquer – start with the base cases and build up to the desired solution, storing results to avoid redundant computation  usually more effective than top-down recursion if the parts are not completely independent (thus leading to redundancy)  note: can implement by adding caching to top-down recursion greedy: involves making a sequence of choices/actions, each of which simply looks best at the moment  applicable when a solution is a sequence of moves & perfect knowledge is available backtracking: involves making a sequence of choices/actions (similar to greedy), but stores alternatives so that they can be attempted if the current choices lead to failure  more costly in terms of time and memory than greedy, but general-purpose  worst case: will have to try every alternative and exhaust the search space